The full design of the sky-skills harness — nine generator skills and two evaluators, now nine components (a self-evolution loop, 09, was added). Eight are done; the canonical library (02) just reached the full 59/59 matrix (verified by bin/design-review --coverage); and design-evolve (09) is built but has not run an evolution round yet. /design-loop (06) shipped its first real run — an anthropic security page at critic 92, now the first page in the corpus. Every component must defend one question: what specific model weakness does it correspond to? No weakness, no seat at the table.
sky-skills 架构的完整蓝图 —— 九个 生成器 skill 加两个 评审员，现在是九个组件(新增了自进化循环 09)。八个已完成；canonical 库(02)刚补满 59/59 整张覆盖表(用 bin/design-review --coverage 实测)；design-evolve(09)机制已建、但还没跑过一轮进化。/design-loop(06)完成首跑——一个 anthropic security 页 critic 92，已是 corpus 第一页。每个组件都必须能回答一个问题：它挡的是模型的哪一个具体弱点？挡不住弱点，就不配留。
GAN (Goodfellow et al., 2014) — a generator and a discriminator train adversarially. The sharper the discriminator, the stronger the generator becomes. Our nine design skills are the generator, design-review is the discriminator.
GAN(Goodfellow et al., 2014) —— generator 和 discriminator 对抗训练。discriminator 越锐利,generator 就越强。我们的 9 个设计 skill 是 generator,design-review 是 discriminator。
Anthropic, Harness design for long-running apps — planner / generator / evaluator three-part split, evaluators must be separate, sprint contracts, 5–15 iteration rounds, and the rule we take most seriously: each harness component must correspond to one specific model weakness. Assumptions worth stress-testing, because they may be incorrect or quickly outdated.
Anthropic 的 Harness design for long-running apps —— planner / generator / evaluator 三段式,evaluator 必须独立,sprint contract,5–15 轮迭代。我们最认真对待的一条:每一个 harness 组件都要对应模型的一个具体弱点。这些假设值得压测,因为它们可能不对,或者很快过时。
Agents tend to praise their own work confidently — even when quality is obviously mediocre to a human observer. Separating the agent doing the work from the agent evaluating it is a powerful lever. Agent 倾向于自信地称赞自己的作品 —— 哪怕在人类观察者看来质量明显平庸。把做事的 agent 和评判的 agent 分开，是一个真正的杠杆。
Anthropic — Harness design postAnthropic —— harness-design 文章
A one-sentence brief becomes a web page that stands next to anthropic.com, stripe.com, linear.app. Not technically bug-free
— taste, originality, rhythm, copy, hand-crafted illustrations, device-agnostic, accessible, shareable.
一句话需求，产出能和 anthropic.com / stripe.com / linear.app 同台站着的一张页面。不是"技术上没 bug"，而是 有品位、有原创、有节奏、copy 过关、图示手工感强、全设备可用、无障碍、可分享。
The only legitimate reason for a harness component to exist is that a specific model weakness makes it necessary. If a reader can argue this weakness isn't real
and we can't produce evidence, we remove the component.
一个 harness 组件能存在的唯一理由，是有一条具体的模型弱点要求它存在。如果有人说"这条弱点不成立"，而我们拿不出证据，就把这个组件删掉。
skills/design-review/scripts/sprint-contract.mjs + bin/design-review --plan + skills/design-planner/SKILL.md
Model weakness对应弱点
Faced with a vague brief, the model fills in training-data medians. Scope drifts. A planner expands the brief into a sprint contract — section list, must-include content, hard constraints — before the generator runs. 面对模糊需求，模型会用训练数据的中位数填空，scope 失控。planner 在生成器动手前，把需求展开成 sprint contract —— 必含 section、必含内容、硬指标。
skills/<style>-design/references/canonical/
Model weakness对应弱点
Without reference, output regresses to the most average web design.
A per-style canonical library of 10 page-types (pricing / landing / docs-home / blog / product / team / comparison / FAQ / changelog / feature-deep) anchors quality at each style's own best — not an industry average.
没有参考，输出就向"最平均的 web 设计"收敛。每个 style 自己的 10 个 page-type canonical 库(pricing / landing / docs / blog / product / team / comparison / FAQ / changelog / feature-deep)，把品质锚定在该 style 自己的最好，不是行业平均。
skills/{apple,anthropic,ember,sage}-design/
Model weakness对应弱点
Models don't naturally articulate their design decisions after writing. Without a required self-diff note, the critic has no concrete target — just vibes. Enforced 2026-04-24 via cross-skill-rules §M + verify.py canonical-self-diff check: every canonical HTML must embed a design-review:self-diff v1 comment block (5+ decisions × [id] chose A over B + known trade-offs). All 14 existing canonicals back-filled from their .md files.
模型写完 HTML 不会自然记录设计决策。没有强制的 self-diff note，critic 没有具体靶子，只能凭感觉评。2026-04-24 正式强制：cross-skill-rules §M + verify.py 的 canonical-self-diff check：每张 canonical 必须 embed design-review:self-diff v1 注释块(5+ 条 [id] chose A over B 决策 + known trade-offs)。14 张现有 canonical 从各自 .md 全部回补完毕。
skills/design-review/
Model weakness对应弱点
Models don't self-check runtime behavior — multi-viewport rendering, keyboard navigation, color contrast, LCP/CLS, SEO meta. A mechanical gate catches the class of bugs source-scanning and model self-review both miss. runtime 行为(多视口渲染、键盘可达性、对比度、LCP/CLS、SEO meta)模型自己不会检查。mechanical 检查抓的正是源码扫描和模型自评都抓不到的那一类 bug。
.claude/agents/design-{composition,copy,illustration,brand}-critic.md
Model weakness对应弱点
Evaluating one's own work tends to produce uniform praise. Even a fresh-context critic has a limited attention budget — one reviewer can't divide it evenly across four orthogonal axes: composition, copy, illustration, brand. Four specialists (weights 25/25/20/30) in fresh contexts + an aggregator. First live run on this page: solo-critic 93, multi-critic 88 — the 5-point gap was the illustration specialist catching a 6th-hue leak inside an SVG the generalist missed. 评审自己的作品会整体说好话。即便 fresh context 的 critic，注意力预算也有限，一个人很难把它均匀分给四个正交维度：版式、copy、插画、品牌。四位专家(权重 25/25/20/30)各自 fresh context + 一位汇总。本页首次实战：单 critic 93，多 critic 88 —— 差的 5 分是插画专家在 SVG 里抓到了通才漏掉的 6 色偷渡。
.claude/commands/design-loop.md
Model weakness对应弱点
Models default to ship the first-pass output.
Without orchestration, actual iteration ends after 1–2 rounds. A loop controller forces the planner → generator → review → critic cycle for up to 5 rounds, or escalates to a human.
模型默认"第一版够用就发"。没有外部编排，实际只跑 1-2 轮。loop controller 强制 planner → generator → review → critic 最多 5 轮，超过还过不去就交还给人。
.claude/agents/design-learner.md + scripts/learning-loop.mjs
Model weakness对应弱点
Models don't automatically archive "the new issue the critic caught today." Without this loop, the same bug class keeps being caught by humans. design-learner reads critic verdicts (solo or multi) and proposes three codifications: new rows in known-bugs.md, new mechanical checks in visual-audit.mjs, per-skill dos-and-donts entries. First live run 2026-04-22: 13 raw verdict issues → 5 unique bug classes → 3 new known-bugs (1.17/1.18/1.19) + 1 new visual-audit check (figure-no-caption) + 8 dos-and-donts rows. Human-in-the-loop — agent proposes; human applies.
模型不会自动归档"critic 今天新抓到的问题"。没有这个环节，同一类 bug 会反复被人抓到。design-learner 读 critic 的 verdict(单/多)，提三件事：known-bugs.md 新行、visual-audit.mjs 新机器 check、每 skill 的 dos-and-donts。首次实战 2026-04-22：13 条原始 issue → 5 类 bug → 新 3 行 known-bugs(1.17/1.18/1.19)+ 1 个新机器 check(figure-no-caption)+ 8 条 dos-and-donts。human-in-the-loop —— agent 提，人审完 apply。
skills/design-review/scripts/library-grower.mjs
Model weakness对应弱点
Models can't self-organize yesterday's good output
into tomorrow's reference. Without this, the canonical library stays fixed and becomes stale over 6 months. Five successful outputs of a page-type → auto-distilled into a new canonical (candidate.html + diffs.md), reviewed by a human, merged. /design-loop deposits each ≥88 ship into the corpus via --deposit.
模型没有"把昨天的好产物整理成明天参考"的能力。没有它，canonical library 会固定，6 个月后就开始过期。同一 page-type 攒 5 次优秀产物 → 自动蒸馏成新 canonical(candidate.html + diffs.md)，人审核通过后并入。/design-loop 每次 ≥88 上线都用 --deposit 把页喂进 corpus。
skills/design-evolve/SKILL.md
Model weakness对应弱点
Models repeat the same taste mistakes and can't tell which of their own rules actually help. A Darwin / autoresearch ratchet: mutate one generator rule, regenerate, score with the frozen evaluator, keep it only if it strictly beats the baseline with no held-out canonical regressing. Selection, not authorship — the system improves its own rules without being able to cheat the test. EvolveR scoring prunes dead rules; a held-out regression gate blocks overfit. 模型会重复同样的审美错误，也分不清自己哪条规则真有用。一个 Darwin / autoresearch 式爬坡：改一条生成器规则、重新生成、用冻结的评审器打分，只有严格超过基线且无 held-out canonical 变差才保留。由选择而非作者决定 —— 系统能改进自己的规则，又没法在测试上作弊。EvolveR 成功率裁剪失效规则；held-out 回退门挡过拟合。
The pipeline is built end-to-end, in the order hard dependencies allowed. What's left isn't new machinery — the canonical library (02) just reached its full 59/59 matrix, so all that remains is one first run (09, design-evolve, yet to attempt an evolution round). 管线已端到端建好，顺序由硬依赖决定。剩下的不是新机器 —— canonical 库(02)刚补满 59/59 整张覆盖表，只剩一次首跑(09 design-evolve 还没试过一轮进化)。
The most dangerous failure mode is that as we build more shared layers, the 9 styles start producing outputs that all look the same.
The rule below makes that mechanically impossible: universal rules never judge style; style rules live in per-style folders; critic's taste is anchored to each style's own canonical.
最危险的失败模式：共享层越多，9 个风格的产出越趋同。下面这条规矩让这件事在机制上不可能发生 —— universal 规则从不评判风格，风格规则住在各自的文件夹，critic 的品味锚定在每个 style 自己的 canonical。
Eight components done (01 / 02 / 03 / 04 / 05 / 06 / 07 / 08); 02 (canonical library) reached the full 59/59 matrix. (06) /design-loop shipped its first real end-to-end run — an anthropic security page at critic 92, now the first page in the corpus. (09) design-evolve — the Darwin / autoresearch self-evolution ratchet (mutate one rule, regenerate, score with the frozen evaluator, keep only if measurably better and no held-out canonical regresses) — is built but has not run an evolution round yet: rules.json is empty and the ledger holds only the loop's first-run deposit. (04) mechanical critic — 38-check visual-audit.mjs (brand-presence / italic-overuse / cross-skill-smell / hollow-card / asymmetric-first-col-hero / svg-foreign-hex / figure-no-caption, via pngjs) + 94 known-bugs catalogued; (01) sprint-contract — planner phase 1 — shipped as sprint-contract.mjs + bin/design-review --plan, contract schema still growing; skills/design-planner/SKILL.md (brief→contract guidance) + LOW-CONFIDENCE fallback for canonical-less page-types shipped 2026-06-11; (05) multi-critic — 4 critic subagents in parallel (composition / copy / illustration / brand, weighted 25/25/20/30) + solo design-critic; (07) learning-loop — design-learner reads verdicts and proposes known-bugs rows + new mechanical checks. 59/59 canonical page-types covered — the full matrix — 10 page-types × anthropic / apple / ember / sage, 5 for atelier, 4 for glass, 3 each for eclat, lectern and primer (a keynote has no pricing page, a board deck has no landing, an application has no hero; the matrix is ragged by design); matrix complete. A second independent evaluator ships as skill 11: gated-dual-clone-audit (4 tiers, paired with skill 10 generator) — same generator/evaluator split principle, different domain. Dated increments since 2026-04-27 — v2/v3 features, new gates, new skills — live in the changelog below.
已完成 8 个组件（01 / 02 / 03 / 04 / 05 / 06 / 07 / 08）；02(canonical 库)已补满 59/59 整张覆盖表。(06) /design-loop 完成首次端到端实战 —— 一个 anthropic security 页 critic 92，已是 corpus 第一页。(09) design-evolve —— Darwin / autoresearch 式自进化爬坡（改一条规则、重新生成、用冻结评审器打分，只有可度量更优且无 held-out canonical 变差才保留）—— 机制已建、但还没跑过一轮进化：rules.json 为空，ledger 里只有 loop 首跑的落库记录。(04) 机器 critic —— 51 项 visual-audit.mjs(brand-presence / italic-overuse / cross-skill-smell / hollow-card / asymmetric-first-col-hero / badge-not-referenced / series-viewbox-drift，用 pngjs)+ 94 条 known-bugs；(01) sprint-contract —— planner 期 1，已发（sprint-contract.mjs + bin/design-review --plan），contract 字段还在扩；skills/design-planner/SKILL.md（brief→contract 指南）+ 无 canonical 类型的 LOW-CONFIDENCE 回退已于 2026-06-11 上线；(05) multi-critic —— 4 位 critic subagent 并行(版式 / 文案 / 插画 / 品牌，权重 25/25/20/30)+ 单 design-critic；(07) learning-loop —— design-learner 读 verdict 自动给出 known-bugs 行 + 新机器 check 建议。覆盖 59/59 canonical page-type —— 整张覆盖表 —— 10 个 page-type × anthropic / apple / ember / sage，atelier 6 个，glass 4 个，eclat、lectern、primer 各 3 个（发布会没有定价页，董事会 deck 没有落地页，应用没有 hero；这张表本就是不规则的）；覆盖表铺满。第二个独立 evaluator 作为第 11 个 skill 落地：gated-dual-clone-audit(4 tier，配对第 10 号 generator)—— 相同的 generator/evaluator 分离原则，不同领域。2026-04-27 起的日期增量（v2/v3 特性、新检查、新 skill）见下方 changelog。
Complete — all 59 shipped, atelier (6) and primer (3) included; verify with bin/design-review --coverage. Each page: full .html + binding .md design-decisions file, all passing verify.py + visual-audit.mjs + screenshot.mjs 0-warning and a --multi-critic pass. They were built 1–2 per focused session — never rushed, to protect the library's reference quality.
已完成 —— 全部 59 张交付，含 atelier 6 张与 primer 3 张；实时计数见 bin/design-review --coverage。每张：完整 .html + 绑定的 .md 设计决策文档，全部过 verify.py + visual-audit.mjs + screenshot.mjs 0 warning 并过 --multi-critic。每个专注 session 做 1–2 张，从不赶工，护住 library 的参考水准。
Three components in dependency order, now all built (2026-06-14): (01) planner expands a brief into a sprint-contract — shipped with --format=json so the loop reads acceptance criteria machine-readably. (06) /design-loop orchestrates planner → generator → review → critic for up to 5 rounds — a deterministic driver (design-loop.mjs) now owns the round counter, the ≥88 ship gate, and deposit-on-ship; its first real end-to-end run shipped an anthropic security page at critic 92. (08) library-grower distills 5+ same-type pages into a candidate (candidate.html + diffs.md), fed by the loop's --deposit. And the line now closes with a 9th: (09) design-evolve, which optimizes the generators' own rules against the frozen evaluator — mechanism built 2026-06-14, awaiting its first evolution round.
三个组件按依赖顺序，现在都已建（2026-06-14）：(01) planner 把需求展开成 sprint-contract —— 已加 --format=json，loop 可机读验收标准。(06) /design-loop 编排 planner → generator → review → critic 最多 5 轮 —— 确定性 driver（design-loop.mjs）现在管轮数计数、≥88 ship 门、上线落库；首次端到端实战已通过——anthropic security 页 critic 92。(08) library-grower 把 5+ 张同类页蒸馏成候选（candidate.html + diffs.md），由 loop 的 --deposit 喂数据。这条线现在以第 9 个收尾：(09) design-evolve，对着冻结评审器优化生成器自己的规则 —— 机制已建（2026-06-14），等第一轮进化。
All 23 skills — plus the harness's slash-commands (/design-loop, /design-distill, /gdc-audit-critic) and six critic subagents — now install globally under ~/.claude/. The whole design harness runs from any repository: ask in natural language or type the slash command, and the generator self-reviews through a CWD-independent launcher (~/.claude/skills/design-review/dr-cli) that locates its own engine no matter where it is called. The CLI's self-location was hardened with readlink -f; reference paths now anchor to the install location instead of the repo root; and the contract no longer writes to /tmp (read from stdout, or written beside the output). The engine was already CWD-independent — only the instruction text had assumed you stood inside the repo.
全部 23 个 skill —— 加上 harness 的斜杠命令(/design-loop、/design-distill、/gdc-audit-critic)和六个 critic subagent —— 现在都装到 ~/.claude/ 全局生效。整套设计 harness 在任意仓库都能跑：自然语言提问或敲斜杠命令，生成器通过一个与当前目录无关的入口(~/.claude/skills/design-review/dr-cli)自评审，无论在哪调用都能定位自己的引擎。CLI 自定位用 readlink -f 加固；引用路径锚到安装位置而非仓库根；合约不再写 /tmp(从 stdout 读，或写在输出文件旁)。引擎本来就与目录无关 —— 只是指令文本之前假设你站在仓库里。
diagram-craft v3 — a genealogy of seven kernel diagram types (§15: block / flowchart / timing-waveform / register-bitfield / ring-buffer / pipeline / scheduler-timeline) and constraints split into three layers: aesthetics immutable, quality machine-gated, structure free to customize. Known-bugs 1.28–1.31 codified into 5 new visual-audit.mjs gates; sprint-contract gains a §1b diagram-density rule; template library grown to anthropic 14 / apple 14 pieces; two diagram galleries shipped (anthropic 23 diagrams, apple 14); gold accent corrected cross-skill from #c49464 to #c9913f. Wave 2: visual-audit grows keyboard-a11y + LCP/CLS perf gates (26 checks); verify.py gains a warn tier with SEO meta checks; --audit reports per-kind stats; --distill wired (component 08); design-planner ships as the 13th skill (component 01). Later the same day, diagram-craft reaches the remaining two generators: ember-design and sage-design each gain a diagram-craft.md + 8 diagram templates + an 8-figure gallery, and the diagram-monochrome gate now applies to anthropic / ember / sage (apple stays exempt — grayscale is its identity).
diagram-craft v3 —— 内核七图型谱系（§15：框图 / 流程图 / 时序波形 / 寄存器位域 / 环形缓冲 / 流水线 / 调度时间线）+ 约束三层化（审美不可变 / 质量机器检查 / 结构自由定制）。known-bugs 1.28–1.31 写成 visual-audit.mjs 5 个新检查；sprint-contract 新增 §1b 图密度条款；模板库扩到 anthropic 14 件 / apple 14 件；双 diagram gallery 上线（anthropic 23 图、apple 14 图）；金主色 cross-skill 修正：#c49464 → #c9913f。Wave 2：visual-audit 新增 keyboard 可达性 + LCP/CLS 性能检查（26 项 check）；verify.py 新增 warn 档并带 SEO meta 检查；--audit 输出按 kind 统计；--distill 接线完成（组件 08）；design-planner 作为第 13 个 skill 上线（组件 01）。同日收尾：diagram-craft 补齐剩下两个 generator —— ember-design 与 sage-design 各得 diagram-craft.md + 8 件图示模板 + 8 图 gallery；diagram-monochrome 检查归属定为 anthropic / ember / sage 生效，apple 凭灰阶身份豁免。
anthropic-design v3 — four-piece md rendering pipeline under scripts/: md-mirror (1 .md → 1 anthropic-styled .html with inline CSS), md-rewrite-links (in-place *.md→*.html href swap), md-pack (fold linked .md into a flat _md/ subdir + retarget hrefs + basename rescue for ../ typos in source markdown), cross-link-pack (fold cross-dir sibling .html into the same _md/). Goal: an anthropic-styled doc directory that survives cp -r anywhere with every link intact. Field-tested same day on a 4-HTML / 13-md / 124-link doc set.
anthropic-design v3 —— scripts/ 下 4 件套 md 渲染管线：md-mirror（1 .md → 1 anthropic 风格 .html，内联 CSS）/ md-rewrite-links（原地 *.md→*.html href 替换）/ md-pack（把链到的 .md 折叠到扁平 _md/ 子目录 + 重写 href + basename 救援源文档 ../ typo）/ cross-link-pack（跨目录 sibling .html 也折叠进同款 _md/）。目标：anthropic 风格的文档目录 cp -r 到任何地方所有跳转依然有效。同日实测一份 4-HTML / 13-md / 124-链 的文档集。
anthropic-design v2 — scenario recipes (M1-M10 motion / L1-L10 layout / C1-C11 components) + references/ux-writing.md (CTA / empty / error / placeholder copy + banned-word list) + 11 new recipe classes / 6 keyframes / 4 utility animation classes in assets/anthropic.css. bin/design-review --audit batch mode (file / directory / URL inputs, no exit-on-error, markdown + json reports). 12th skill: doc-review-loop — independent writer/reviewer harness for serious decision documents. Not a design-harness component; it's the iteration pattern of component 06 /design-loop applied to docs — a working proof-of-concept that the loop pattern survives reality.
anthropic-design v2 —— scenario recipes(M1-M10 动效 / L1-L10 版式 / C1-C11 控件)+ references/ux-writing.md(CTA / empty / error / placeholder 文案 + 禁用词清单)+ 11 个新 recipe class / 6 keyframes / 4 utility animation class，全部入 assets/anthropic.css。bin/design-review --audit 批量模式(file / directory / URL 输入，不在错误时 exit，输出 markdown + json 报告)。第 12 个 skill：doc-review-loop —— 给认真决策书的独立 writer/reviewer harness。不是 design-harness 组件，是 组件 06 /design-loop 的迭代模式落到文档场景的实证 —— 证明 loop 模式可以从原理走到生产。
The three voice variants (apple / ember / sage) are updated in the same PR as this document. 三个 voice 变体（apple / ember / sage）随主文档同一个 PR 更新。
known-bugs.md and add a check.known-bugs.md + 加一个 checkSimplicity firstis a heuristic, not a hard rule. "simplicity first" 是经验法则，不是硬规则。
While designing this roadmap, I (Claude) once argued for cutting 8 components to 5, citing the article's simplicity first
principle. The user correctly pushed back: Why not 8 again?
The article's actual hard rule isn't fewer is better
— it's every component must correspond to a specific model weakness.
The retreat was unwarranted and I restored all 8. Lesson: when tempted to cut a component, the only valid argument is evidence that the weakness it defends doesn't exist. Heuristic shortcuts are how real weaknesses go undefended.
写这份 roadmap 时，我(Claude)曾经引用文章里 "simplicity first" 把 8 组件砍到 5。用户正确地反问"为什么又不是 8 了"。文章真正的硬规矩不是"越少越好"，是"每个组件对应一个具体弱点"。那次退让没有依据，我恢复了 8 个。教训：想裁一个组件时，唯一有效的理由是证据表明它挡的弱点不存在。走捷径裁组件，就是把本该挡住的弱点放开了。
The document rules above apply to outside contributors too. Argue with evidence. We'd rather restructure the harness than keep a component that isn't really doing anything. 上面的文档规矩对外部贡献者同样适用。用证据讲话。我们宁愿重构 harness，也不愿让一个其实没挡任何东西的组件继续占位。